This paper presents a novel optical flow-based framework for monocular visual odometry (VO), combining GMFlow for high-quality optical flow estimation and a ConvLSTM-based pose estimation network. The proposed framework effectively captures both spatial and temporal motion relationships, addressing challenges such as motion estimation, scale ambiguity, and dynamic environments in monocular VO tasks. We evaluated the framework on the KITTI VO/SLAM benchmark dataset. Experimental results demonstrate that the model achieves superior accuracy in both translational and rotational pose estimation compared to classical methods like DeepVO and Flowdometry. The framework exhibits strong adaptability across diverse driving scenarios, including urban streets, highways, and rural roads, while maintaining robust performance under varying sequence lengths and speeds. By leveraging deep learning techniques for feature extraction and temporal modeling, the proposed framework significantly enhances the accuracy and robustness of monocular VO. Future work will explore incorporating attention mechanisms and extending the framework to handle dynamic environments for improved real-world applicability.